4 ms·
People complained that new AI is "stealing" from artists. But stealing from other AI turns out to often be easier. And this is where things get fun, because c
by dathinab 4y ago
People complained that new AI is "stealing" from artists.
But stealing from other AI turns out to often be easier.
And this is where things get fun, because companies like OpenAI want to be able to train on all the data without any explicit permissions from the creators, but the moment people do the same to them they likely (we will see) be very much against it.
So it will be interesting if they will be able to both have and eat the cake (e.g. by using Microsoft lobby to push absurd law) or will they fall apart due to cannibalization making it non profitable to create better AI.
EDIT: This comment isn't specific to Google/Bert, so it doesn't matter weather Google actually did so or weather not.
- commoner 4y agoI can see the GitHub Copilot controversy being resolved in this way. If Microsoft, GitHub, and OpenAI successfully use the fair use defense for Copilot's appropriation of proprietary and incompatibly licensed code, then a free and open source alternative to Copilot can be trained on Copilot's outputs. After all, the GitHub Copilot Product Specific Terms say: > 2. Ownership of Suggestions and Your Code > GitHub does not claim any ownership rights in Suggestions. You retain ownership of Your Code. https://github.com/customer-terms/github-copilot-product-specific-terms https://github.com/customer-terms/github-copilot-product-spe...
- circuit10 4y agoWhy would it need to be trained on Copilot’s output? Its training data is publicly available code on GitHub, so just use that directly. ChatGPT is different because they specifically trained it as an assistant with a private dataset
- commoner 4y agoThat was in reference to GP's point that > stealing from other AI turns out to often be easier If that holds true for Copilot, then it would make sense for a free and open source alternative to train on Copilot's outputs.